Novel Machine Learning Modeling Approach for Fatigue Failure of Hydrogen-Transporting Pipelines

Summary The challenge of data scarcity significantly impedes machine learning (ML) applications in specialized domains, such as research on hydrogen embrittlement in carbon steel, posing a significant obstacle to repurposing existing natural gas networks for hydrogen transportation. Hence, this study aims to develop an ML model to predict hydrogen-assisted fatigue failures in hydrogen-transporting gas pipelines. A novel approach is developed to establish an accurate and widely applicable ML model that leverages data from one set of mechanical tests to improve the prediction of other relevant mechanical parameters. The new approach involves employing data extracted from primary literature sources that conducted the fatigue crack growth (FCG) test. These data are then utilized in an existing ML model, previously trained with tensile test data, to predict the reduction of area (RA) of materials subject to hydrogen embrittlement. Building upon the aforementioned approach, the model-generated RA is further integrated as an input parameter to train a new ML model aimed at predicting FCG rates. In parallel, a conventional ML model is developed that excludes model-calculated RA as an input parameter, focusing solely on fatigue test variables. The traditional approach is undertaken to evaluate the efficacy of incorporating model-generated RA into the model. Data sets from one source (i.e., 96 data points) are completely removed from the database to compare the models. Afterward, model predictions are compared with the removed data set to assess model performance and accuracy. Testing the models with the removed data sets shows the traditional model has a mean relative error (MRE) of 11.65, while the RA-based model has an MRE of 10.06. Moreover, the model with RA is 15.65% more accurate than the traditional model in terms of MRE after data removal. Furthermore, both models are assessed using standard statistical performance parameters such as the coefficient of determination (R2), and they show similar performance. Despite the similarity in conventional performance metrics for both models, the analysis of maximum error and the scatter in the cross-correlation plots of model predictions clearly indicates that the model incorporating the calculated RA demonstrates closer alignment with actual values. This outcome not only underscores the merit of this innovative approach but also contributes significantly to mitigating data scarcity challenges in the evolving field of hydrogen energy. Data scarcity challenges in hydrogen embrittlement studies are addressed by developing a novel approach that uses data from one type of mechanical test to improve the prediction accuracy of a model of other relevant mechanical parameters. The method developed in this study can be extended for other ML modeling applications when data scarcity is the limiting factor.